369 lines
16 KiB
C++
369 lines
16 KiB
C++
// Copyright 2019 Google LLC
|
|
//
|
|
// This source code is licensed under the BSD-style license found in the
|
|
// LICENSE file in the root directory of this source tree.
|
|
|
|
#include <algorithm>
|
|
#include <cfloat>
|
|
#include <cmath>
|
|
#include <cstdint>
|
|
#include <functional>
|
|
#include <random>
|
|
#include <vector>
|
|
|
|
#include "conv.h"
|
|
#include "utils.h"
|
|
#include "xnnpack.h"
|
|
#include "xnnpack/common.h"
|
|
#include "xnnpack/igemm.h"
|
|
#include "xnnpack/indirection.h"
|
|
#include "xnnpack/math.h"
|
|
#include "xnnpack/microfnptr.h"
|
|
#include "xnnpack/microparams-init.h"
|
|
#include "xnnpack/pack.h"
|
|
#include "xnnpack/buffer.h"
|
|
#include <benchmark/benchmark.h>
|
|
|
|
static void f16_igemm(benchmark::State& state,
|
|
xnn_f16_igemm_minmax_ukernel_fn igemm,
|
|
xnn_init_f16_minmax_params_fn init_params,
|
|
uint32_t mr, uint32_t nr, uint32_t kr, uint32_t sr,
|
|
benchmark::utils::IsaCheckFunction isa_check = nullptr)
|
|
{
|
|
if (isa_check != nullptr && !isa_check(state)) {
|
|
return;
|
|
}
|
|
|
|
const size_t input_height = state.range(0);
|
|
const size_t input_width = state.range(1);
|
|
const size_t kernel_height = state.range(2);
|
|
const size_t kernel_width = state.range(3);
|
|
const size_t kernel_size = kernel_height * kernel_width;
|
|
const size_t padding_height = state.range(4);
|
|
const size_t padding_width = state.range(5);
|
|
const size_t subsampling = state.range(6);
|
|
const size_t dilation = state.range(7);
|
|
const size_t group_input_channels = state.range(8);
|
|
const size_t group_output_channels = state.range(9);
|
|
|
|
std::random_device random_device;
|
|
auto rng = std::mt19937(random_device());
|
|
auto f32rng = std::bind(std::uniform_real_distribution<float>(), std::ref(rng));
|
|
|
|
const size_t output_pixel_stride = group_output_channels;
|
|
const size_t input_pixel_stride = group_input_channels;
|
|
const size_t effective_kernel_height = (kernel_height - 1) * dilation + 1;
|
|
const size_t effective_kernel_width = (kernel_width - 1) * dilation + 1;
|
|
const size_t padding_left = padding_width / 2;
|
|
const size_t padding_top = padding_height / 2;
|
|
const size_t output_height = (input_height + padding_height - effective_kernel_height) / subsampling + 1;
|
|
const size_t output_width = (input_width + padding_width - effective_kernel_width) / subsampling + 1;
|
|
const size_t output_size = output_height * output_width;
|
|
|
|
const size_t mc_stride = benchmark::utils::RoundUp<size_t>(output_size, mr);
|
|
const size_t nc_stride = benchmark::utils::RoundUp<size_t>(group_output_channels, nr);
|
|
const size_t kc_stride = benchmark::utils::RoundUp<size_t>(group_input_channels, kr * sr);
|
|
|
|
xnnpack::Buffer<xnn_float16> a(input_height * input_width * input_pixel_stride + XNN_EXTRA_BYTES / sizeof(xnn_float16));
|
|
std::generate(a.begin(), a.end(), f32rng);
|
|
xnnpack::Buffer<xnn_float16> k(group_output_channels * kernel_height * kernel_width * group_input_channels);
|
|
std::generate(k.begin(), k.end(), f32rng);
|
|
xnnpack::Buffer<xnn_float16> b(group_output_channels);
|
|
std::generate(b.begin(), b.end(), f32rng);
|
|
|
|
xnnpack::Buffer<xnn_float16> z(group_input_channels + XNN_EXTRA_BYTES / sizeof(xnn_float16));
|
|
|
|
const size_t w_elements = (kernel_size * kc_stride + 1) * nc_stride;
|
|
const size_t i_elements = mc_stride * kernel_size;
|
|
const size_t c_elements = output_height * output_width * output_pixel_stride;
|
|
const size_t num_buffers = 1 +
|
|
benchmark::utils::DivideRoundUp<size_t>(benchmark::utils::GetMaxCacheSize(),
|
|
sizeof(xnn_float16) * (w_elements + c_elements) + sizeof(void*) * i_elements);
|
|
|
|
xnnpack::Buffer<xnn_float16, XNN_ALLOCATION_ALIGNMENT> w(w_elements * num_buffers);
|
|
xnn_pack_f16_conv_goki_w(/*groups=*/1, group_output_channels, kernel_size, group_input_channels, nr, kr, sr,
|
|
reinterpret_cast<const uint16_t*>(k.data()),
|
|
reinterpret_cast<const uint16_t*>(b.data()),
|
|
/*scale=*/nullptr,
|
|
reinterpret_cast<uint16_t*>(w.data()),
|
|
/*extra_bytes=*/0, /*params=*/nullptr);
|
|
for (size_t n = 1; n < num_buffers; n++) {
|
|
std::copy(w.cbegin(), w.cbegin() + w_elements, w.begin() + n * w_elements);
|
|
}
|
|
|
|
xnnpack::Buffer<const xnn_float16*> i(i_elements * num_buffers);
|
|
const size_t tiled_output_size = round_up(output_size, mr);
|
|
xnn_indirection_init_conv2d(
|
|
/*output_tile_size=*/mr,
|
|
/*output_start=*/0,
|
|
/*output_end=*/tiled_output_size,
|
|
reinterpret_cast<const void**>(i.data()),
|
|
a.data(),
|
|
z.data(),
|
|
input_pixel_stride << XNN_LOG2_SIZEOF_HALF,
|
|
input_height, input_width,
|
|
output_height, output_width,
|
|
kernel_height, kernel_width,
|
|
subsampling, subsampling,
|
|
dilation, dilation,
|
|
padding_top, padding_left);
|
|
for (size_t n = 1; n < num_buffers; n++) {
|
|
std::copy(i.cbegin(), i.cbegin() + i_elements, i.begin() + n * i_elements);
|
|
}
|
|
|
|
xnnpack::Buffer<xnn_float16> c(c_elements * num_buffers);
|
|
|
|
// Prepare minmax parameters.
|
|
xnn_f16_minmax_params params;
|
|
init_params(¶ms, static_cast<xnn_float16>(-INFINITY), static_cast<xnn_float16>(INFINITY));
|
|
|
|
size_t buffer_index = 0;
|
|
for (auto _ : state) {
|
|
state.PauseTiming();
|
|
benchmark::utils::PrefetchToL1(a.data(), a.size() * sizeof(xnn_float16));
|
|
buffer_index = (buffer_index + 1) % num_buffers;
|
|
state.ResumeTiming();
|
|
|
|
for (uint32_t m = 0; m < output_size; m += mr) {
|
|
const uint32_t mb = min(output_size - m, mr);
|
|
for (uint32_t n = 0; n < group_output_channels; n += nr) {
|
|
const uint32_t nb = min(group_output_channels - n, nr);
|
|
igemm(
|
|
mb, nb, group_input_channels * sizeof(xnn_float16), kernel_size * mr * sizeof(void*),
|
|
reinterpret_cast<const xnn_float16**>(i.data()) + buffer_index * i_elements + m,
|
|
w.data() + buffer_index * w_elements + n * (kc_stride * kernel_size + 1),
|
|
c.data() + buffer_index * c_elements + m * group_output_channels + n, group_output_channels * sizeof(xnn_float16), nr * sizeof(xnn_float16),
|
|
0, z.data(), ¶ms);
|
|
}
|
|
}
|
|
}
|
|
|
|
const uint64_t cpu_frequency = benchmark::utils::GetCurrentCpuFrequency();
|
|
if (cpu_frequency != 0) {
|
|
state.counters["cpufreq"] = cpu_frequency;
|
|
}
|
|
|
|
state.counters["FLOPS"] = benchmark::Counter(
|
|
uint64_t(state.iterations()) * 2 *
|
|
output_height * output_width *
|
|
group_input_channels * group_output_channels *
|
|
kernel_height * kernel_width,
|
|
benchmark::Counter::kIsRate);
|
|
}
|
|
|
|
#if XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY
|
|
static void f16_igemm_6x16__asm_aarch64_neonfp16arith_cortex_a55(benchmark::State& state, const char* net) {
|
|
f16_igemm(state,
|
|
xnn_f16_igemm_minmax_ukernel_6x16__asm_aarch64_neonfp16arith_cortex_a55,
|
|
xnn_init_f16_minmax_scalar_params,
|
|
/*mr=*/6, /*nr=*/16, /*kr=*/1, /*sr=*/1,
|
|
benchmark::utils::CheckNEONFP16ARITH);
|
|
}
|
|
static void f16_igemm_6x16__asm_aarch64_neonfp16arith_cortex_a55r0(benchmark::State& state, const char* net) {
|
|
f16_igemm(state,
|
|
xnn_f16_igemm_minmax_ukernel_6x16__asm_aarch64_neonfp16arith_cortex_a55r0,
|
|
xnn_init_f16_minmax_scalar_params,
|
|
/*mr=*/6, /*nr=*/16, /*kr=*/1, /*sr=*/1,
|
|
benchmark::utils::CheckNEONFP16ARITH);
|
|
}
|
|
static void f16_igemm_6x16__asm_aarch64_neonfp16arith_cortex_a75(benchmark::State& state, const char* net) {
|
|
f16_igemm(state,
|
|
xnn_f16_igemm_minmax_ukernel_6x16__asm_aarch64_neonfp16arith_cortex_a75,
|
|
xnn_init_f16_minmax_scalar_params,
|
|
/*mr=*/6, /*nr=*/16, /*kr=*/1, /*sr=*/1,
|
|
benchmark::utils::CheckNEONFP16ARITH);
|
|
}
|
|
static void f16_igemm_6x16__asm_aarch64_neonfp16arith_ld64(benchmark::State& state, const char* net) {
|
|
f16_igemm(state,
|
|
xnn_f16_igemm_minmax_ukernel_6x16__asm_aarch64_neonfp16arith_ld64,
|
|
xnn_init_f16_minmax_scalar_params,
|
|
/*mr=*/6, /*nr=*/16, /*kr=*/1, /*sr=*/1,
|
|
benchmark::utils::CheckNEONFP16ARITH);
|
|
}
|
|
static void f16_igemm_4x16__asm_aarch64_neonfp16arith_ld32(benchmark::State& state, const char* net) {
|
|
f16_igemm(state,
|
|
xnn_f16_igemm_minmax_ukernel_4x16__asm_aarch64_neonfp16arith_ld32,
|
|
xnn_init_f16_minmax_scalar_params,
|
|
/*mr=*/4, /*nr=*/16, /*kr=*/1, /*sr=*/1,
|
|
benchmark::utils::CheckNEONFP16ARITH);
|
|
}
|
|
static void f16_igemm_4x16__asm_aarch64_neonfp16arith_ld64(benchmark::State& state, const char* net) {
|
|
f16_igemm(state,
|
|
xnn_f16_igemm_minmax_ukernel_4x16__asm_aarch64_neonfp16arith_ld64,
|
|
xnn_init_f16_minmax_scalar_params,
|
|
/*mr=*/4, /*nr=*/16, /*kr=*/1, /*sr=*/1,
|
|
benchmark::utils::CheckNEONFP16ARITH);
|
|
}
|
|
static void f16_igemm_1x16__asm_aarch64_neonfp16arith_ld32(benchmark::State& state, const char* net) {
|
|
f16_igemm(state,
|
|
xnn_f16_igemm_minmax_ukernel_1x16__asm_aarch64_neonfp16arith_ld32,
|
|
xnn_init_f16_minmax_scalar_params,
|
|
/*mr=*/1, /*nr=*/16, /*kr=*/1, /*sr=*/1,
|
|
benchmark::utils::CheckNEONFP16ARITH);
|
|
}
|
|
static void f16_igemm_1x16__asm_aarch64_neonfp16arith_ld64(benchmark::State& state, const char* net) {
|
|
f16_igemm(state,
|
|
xnn_f16_igemm_minmax_ukernel_1x16__asm_aarch64_neonfp16arith_ld64,
|
|
xnn_init_f16_minmax_scalar_params,
|
|
/*mr=*/1, /*nr=*/16, /*kr=*/1, /*sr=*/1,
|
|
benchmark::utils::CheckNEONFP16ARITH);
|
|
}
|
|
|
|
BENCHMARK_CONV(f16_igemm_6x16__asm_aarch64_neonfp16arith_cortex_a55)
|
|
BENCHMARK_CONV(f16_igemm_6x16__asm_aarch64_neonfp16arith_cortex_a55r0)
|
|
BENCHMARK_CONV(f16_igemm_6x16__asm_aarch64_neonfp16arith_cortex_a75)
|
|
BENCHMARK_CONV(f16_igemm_6x16__asm_aarch64_neonfp16arith_ld64)
|
|
BENCHMARK_CONV(f16_igemm_4x16__asm_aarch64_neonfp16arith_ld32)
|
|
BENCHMARK_CONV(f16_igemm_4x16__asm_aarch64_neonfp16arith_ld64)
|
|
BENCHMARK_CONV(f16_igemm_1x16__asm_aarch64_neonfp16arith_ld32)
|
|
BENCHMARK_CONV(f16_igemm_1x16__asm_aarch64_neonfp16arith_ld64)
|
|
#endif // XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY
|
|
|
|
#if XNN_ENABLE_ARM_FP16_VECTOR && (XNN_ARCH_ARM || XNN_ARCH_ARM64)
|
|
static void f16_igemm_1x8__neonfp16arith_ld64(benchmark::State& state, const char* net) {
|
|
f16_igemm(state,
|
|
xnn_f16_igemm_minmax_ukernel_1x8__neonfp16arith_ld64,
|
|
xnn_init_f16_minmax_scalar_params,
|
|
/*mr=*/1, /*nr=*/8, /*kr=*/1, /*sr=*/1,
|
|
benchmark::utils::CheckNEONFP16ARITH);
|
|
}
|
|
static void f16_igemm_4x8__neonfp16arith_ld64(benchmark::State& state, const char* net) {
|
|
f16_igemm(state,
|
|
xnn_f16_igemm_minmax_ukernel_4x8__neonfp16arith_ld64,
|
|
xnn_init_f16_minmax_scalar_params,
|
|
/*mr=*/4, /*nr=*/8, /*kr=*/1, /*sr=*/1,
|
|
benchmark::utils::CheckNEONFP16ARITH);
|
|
}
|
|
static void f16_igemm_6x8__neonfp16arith_ld64(benchmark::State& state, const char* net) {
|
|
f16_igemm(state,
|
|
xnn_f16_igemm_minmax_ukernel_6x8__neonfp16arith_ld64,
|
|
xnn_init_f16_minmax_scalar_params,
|
|
/*mr=*/6, /*nr=*/8, /*kr=*/1, /*sr=*/1,
|
|
benchmark::utils::CheckNEONFP16ARITH);
|
|
}
|
|
static void f16_igemm_8x8__neonfp16arith_ld64(benchmark::State& state, const char* net) {
|
|
f16_igemm(state,
|
|
xnn_f16_igemm_minmax_ukernel_8x8__neonfp16arith_ld64,
|
|
xnn_init_f16_minmax_scalar_params,
|
|
/*mr=*/8, /*nr=*/8, /*kr=*/1, /*sr=*/1,
|
|
benchmark::utils::CheckNEONFP16ARITH);
|
|
}
|
|
static void f16_igemm_1x16__neonfp16arith_ld64(benchmark::State& state, const char* net) {
|
|
f16_igemm(state,
|
|
xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64,
|
|
xnn_init_f16_minmax_scalar_params,
|
|
/*mr=*/1, /*nr=*/16, /*kr=*/1, /*sr=*/1,
|
|
benchmark::utils::CheckNEONFP16ARITH);
|
|
}
|
|
static void f16_igemm_4x16__neonfp16arith_ld64(benchmark::State& state, const char* net) {
|
|
f16_igemm(state,
|
|
xnn_f16_igemm_minmax_ukernel_4x16__neonfp16arith_ld64,
|
|
xnn_init_f16_minmax_scalar_params,
|
|
/*mr=*/4, /*nr=*/16, /*kr=*/1, /*sr=*/1,
|
|
benchmark::utils::CheckNEONFP16ARITH);
|
|
}
|
|
static void f16_igemm_6x16__neonfp16arith_ld64(benchmark::State& state, const char* net) {
|
|
f16_igemm(state,
|
|
xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64,
|
|
xnn_init_f16_minmax_scalar_params,
|
|
/*mr=*/6, /*nr=*/16, /*kr=*/1, /*sr=*/1,
|
|
benchmark::utils::CheckNEONFP16ARITH);
|
|
}
|
|
static void f16_igemm_8x16__neonfp16arith_ld64(benchmark::State& state, const char* net) {
|
|
f16_igemm(state,
|
|
xnn_f16_igemm_minmax_ukernel_8x16__neonfp16arith_ld64,
|
|
xnn_init_f16_minmax_scalar_params,
|
|
/*mr=*/8, /*nr=*/16, /*kr=*/1, /*sr=*/1,
|
|
benchmark::utils::CheckNEONFP16ARITH);
|
|
}
|
|
|
|
BENCHMARK_CONV(f16_igemm_1x8__neonfp16arith_ld64)
|
|
BENCHMARK_CONV(f16_igemm_4x8__neonfp16arith_ld64)
|
|
BENCHMARK_CONV(f16_igemm_6x8__neonfp16arith_ld64)
|
|
BENCHMARK_CONV(f16_igemm_8x8__neonfp16arith_ld64)
|
|
BENCHMARK_CONV(f16_igemm_1x16__neonfp16arith_ld64)
|
|
BENCHMARK_CONV(f16_igemm_4x16__neonfp16arith_ld64)
|
|
BENCHMARK_CONV(f16_igemm_6x16__neonfp16arith_ld64)
|
|
BENCHMARK_CONV(f16_igemm_8x16__neonfp16arith_ld64)
|
|
#endif // XNN_ENABLE_ARM_FP16_VECTOR && (XNN_ARCH_ARM || XNN_ARCH_ARM64)
|
|
|
|
#if XNN_ARCH_X86 || XNN_ARCH_X86_64
|
|
static void f16_igemm_1x8__avx2_broadcast(benchmark::State& state, const char* net) {
|
|
f16_igemm(state,
|
|
xnn_f16_igemm_minmax_ukernel_1x8__avx2_broadcast,
|
|
xnn_init_f16_minmax_scalar_params,
|
|
/*mr=*/1, /*nr=*/8, /*kr=*/1, /*sr=*/1,
|
|
benchmark::utils::CheckAVX2);
|
|
}
|
|
static void f16_igemm_4x8__avx2_broadcast(benchmark::State& state, const char* net) {
|
|
f16_igemm(state,
|
|
xnn_f16_igemm_minmax_ukernel_4x8__avx2_broadcast,
|
|
xnn_init_f16_minmax_scalar_params,
|
|
/*mr=*/4, /*nr=*/8, /*kr=*/1, /*sr=*/1,
|
|
benchmark::utils::CheckAVX2);
|
|
}
|
|
static void f16_igemm_5x8__avx2_broadcast(benchmark::State& state, const char* net) {
|
|
f16_igemm(state,
|
|
xnn_f16_igemm_minmax_ukernel_5x8__avx2_broadcast,
|
|
xnn_init_f16_minmax_scalar_params,
|
|
/*mr=*/5, /*nr=*/8, /*kr=*/1, /*sr=*/1,
|
|
benchmark::utils::CheckAVX2);
|
|
}
|
|
static void f16_igemm_6x8__avx2_broadcast(benchmark::State& state, const char* net) {
|
|
f16_igemm(state,
|
|
xnn_f16_igemm_minmax_ukernel_6x8__avx2_broadcast,
|
|
xnn_init_f16_minmax_scalar_params,
|
|
/*mr=*/6, /*nr=*/8, /*kr=*/1, /*sr=*/1,
|
|
benchmark::utils::CheckAVX2);
|
|
}
|
|
static void f16_igemm_7x8__avx2_broadcast(benchmark::State& state, const char* net) {
|
|
f16_igemm(state,
|
|
xnn_f16_igemm_minmax_ukernel_7x8__avx2_broadcast,
|
|
xnn_init_f16_minmax_scalar_params,
|
|
/*mr=*/7, /*nr=*/8, /*kr=*/1, /*sr=*/1,
|
|
benchmark::utils::CheckAVX2);
|
|
}
|
|
static void f16_igemm_1x16__avx2_broadcast(benchmark::State& state, const char* net) {
|
|
f16_igemm(state,
|
|
xnn_f16_igemm_minmax_ukernel_1x16__avx2_broadcast,
|
|
xnn_init_f16_minmax_scalar_params,
|
|
/*mr=*/1, /*nr=*/16, /*kr=*/1, /*sr=*/1,
|
|
benchmark::utils::CheckAVX2);
|
|
}
|
|
static void f16_igemm_3x16__avx2_broadcast(benchmark::State& state, const char* net) {
|
|
f16_igemm(state,
|
|
xnn_f16_igemm_minmax_ukernel_3x16__avx2_broadcast,
|
|
xnn_init_f16_minmax_scalar_params,
|
|
/*mr=*/3, /*nr=*/16, /*kr=*/1, /*sr=*/1,
|
|
benchmark::utils::CheckAVX2);
|
|
}
|
|
static void f16_igemm_4x16__avx2_broadcast(benchmark::State& state, const char* net) {
|
|
f16_igemm(state,
|
|
xnn_f16_igemm_minmax_ukernel_4x16__avx2_broadcast,
|
|
xnn_init_f16_minmax_scalar_params,
|
|
/*mr=*/4, /*nr=*/16, /*kr=*/1, /*sr=*/1,
|
|
benchmark::utils::CheckAVX2);
|
|
}
|
|
static void f16_igemm_5x16__avx2_broadcast(benchmark::State& state, const char* net) {
|
|
f16_igemm(state,
|
|
xnn_f16_igemm_minmax_ukernel_5x16__avx2_broadcast,
|
|
xnn_init_f16_minmax_scalar_params,
|
|
/*mr=*/5, /*nr=*/16, /*kr=*/1, /*sr=*/1,
|
|
benchmark::utils::CheckAVX2);
|
|
}
|
|
|
|
BENCHMARK_CONV(f16_igemm_1x8__avx2_broadcast)
|
|
BENCHMARK_CONV(f16_igemm_4x8__avx2_broadcast)
|
|
BENCHMARK_CONV(f16_igemm_5x8__avx2_broadcast)
|
|
BENCHMARK_CONV(f16_igemm_6x8__avx2_broadcast)
|
|
BENCHMARK_CONV(f16_igemm_7x8__avx2_broadcast)
|
|
BENCHMARK_CONV(f16_igemm_1x16__avx2_broadcast)
|
|
BENCHMARK_CONV(f16_igemm_3x16__avx2_broadcast)
|
|
BENCHMARK_CONV(f16_igemm_4x16__avx2_broadcast)
|
|
BENCHMARK_CONV(f16_igemm_5x16__avx2_broadcast)
|
|
#endif // XNN_ARCH_X86 || XNN_ARCH_X86_64
|
|
|
|
#ifndef XNNPACK_BENCHMARK_NO_MAIN
|
|
BENCHMARK_MAIN();
|
|
#endif
|